diff --git a/accel/tcg/atomic_template.h b/accel/tcg/atomic_template.h
index afa8a9daf3..d347462af5 100644
--- a/accel/tcg/atomic_template.h
+++ b/accel/tcg/atomic_template.h
@@ -28,8 +28,8 @@
 # define SHIFT      4
 #elif DATA_SIZE == 8
 # define SUFFIX     q
-# define DATA_TYPE  uint64_t
-# define SDATA_TYPE int64_t
+# define DATA_TYPE  aligned_uint64_t
+# define SDATA_TYPE aligned_int64_t
 # define BSWAP      bswap64
 # define SHIFT      3
 #elif DATA_SIZE == 4
diff --git a/include/qemu/atomic.h b/include/qemu/atomic.h
index fe5467d193..112a29910b 100644
--- a/include/qemu/atomic.h
+++ b/include/qemu/atomic.h
@@ -271,7 +271,19 @@
     _oldn;                                                              \
 })
 
-/* Abstractions to access atomically (i.e. "once") i64/u64 variables */
+/*
+ * Abstractions to access atomically (i.e. "once") i64/u64 variables.
+ *
+ * The i386 abi is odd in that by default members are only aligned to
+ * 4 bytes, which means that 8-byte types can wind up mis-aligned.
+ * Clang will then warn about this, and emit a call into libatomic.
+ *
+ * Use of these types in structures when they will be used with atomic
+ * operations can avoid this.
+ */
+typedef int64_t aligned_int64_t __attribute__((aligned(8)));
+typedef uint64_t aligned_uint64_t __attribute__((aligned(8)));
+
 #ifdef CONFIG_ATOMIC64
 /* Use __nocheck because sizeof(void *) might be < sizeof(u64) */
 #define qatomic_read_i64(P) \
diff --git a/include/qemu/stats64.h b/include/qemu/stats64.h
index fdd3d1b8f9..802402254b 100644
--- a/include/qemu/stats64.h
+++ b/include/qemu/stats64.h
@@ -21,7 +21,7 @@
 
 typedef struct Stat64 {
 #ifdef CONFIG_ATOMIC64
-    uint64_t value;
+    aligned_uint64_t value;
 #else
     uint32_t low, high;
     uint32_t lock;
diff --git a/linux-user/hppa/cpu_loop.c b/linux-user/hppa/cpu_loop.c
index 3aaaf3337c..82d8183821 100644
--- a/linux-user/hppa/cpu_loop.c
+++ b/linux-user/hppa/cpu_loop.c
@@ -82,7 +82,7 @@ static abi_ulong hppa_lws(CPUHPPAState *env)
                 o64 = *(uint64_t *)g2h(cs, old);
                 n64 = *(uint64_t *)g2h(cs, new);
 #ifdef CONFIG_ATOMIC64
-                r64 = qatomic_cmpxchg__nocheck((uint64_t *)g2h(cs, addr),
+                r64 = qatomic_cmpxchg__nocheck((aligned_uint64_t *)g2h(cs, addr),
                                                o64, n64);
                 ret = r64 != o64;
 #else
diff --git a/softmmu/timers-state.h b/softmmu/timers-state.h
index 8c262ce139..94bb7394c5 100644
--- a/softmmu/timers-state.h
+++ b/softmmu/timers-state.h
@@ -47,7 +47,7 @@ typedef struct TimersState {
     int64_t last_delta;
 
     /* Compensate for varying guest execution speed.  */
-    int64_t qemu_icount_bias;
+    aligned_int64_t qemu_icount_bias;
 
     int64_t vm_clock_warp_start;
     int64_t cpu_clock_offset;
diff --git a/util/qsp.c b/util/qsp.c
index bacc5fa2f6..8562b14a87 100644
--- a/util/qsp.c
+++ b/util/qsp.c
@@ -83,8 +83,8 @@ typedef struct QSPCallSite QSPCallSite;
 struct QSPEntry {
     void *thread_ptr;
     const QSPCallSite *callsite;
-    uint64_t n_acqs;
-    uint64_t ns;
+    aligned_uint64_t n_acqs;
+    aligned_uint64_t ns;
     unsigned int n_objs; /* count of coalesced objs; only used for reporting */
 };
 typedef struct QSPEntry QSPEntry;